Post on 06-Sep-2018
Paper to be presented at
DRUID15, Rome, June 15-17, 2015
(Coorganized with LUISS)
The Productivity Effects of Excess Labour Turnover in Young FirmsMartin Murmann
Centre for European Economic ResearchIndustrial Economicsmurmann@zew.de
AbstractThis paper assesses the effects of worker replacement (excess labour turnover) on theproductivity of young firms. Young firms are prone to most risk factors for excessively high turnover rates and are notcovered by most of the existing literature on effects of labour turnover on firm performance. Existing research on theeffects of worker replacement on the performance of established firms is difficult to relate to young firms and showsambiguous results. Using structural identification to account for potential endogeneity of worker replacement, this paperfinds that worker replacement has negative effects on the productivity of young firms. It is disproportionally deleterious in igh-tech manufacturing startups, startups which employ highly educated employees, and in young firms above themedian size and age of the sample. A robustness check indicates that these results are driven by firms with a high shareof quits (by employees) on total separations. Replacements in the sequel of dismissals, initiated by young firmsthemselves, seem to be less detrimental.
Jelcodes:L25,J63
The Productivity Effects of Excess Labour Turnover in
Young Firms
February 28, 2015
Preliminary version - please do not cite or circle
Abstract
This paper assesses the effects of worker replacement (excess labour turnover) on the
productivity of young firms. Young firms are prone to most risk factors for excessively high
turnover rates and are not covered by most of the existing literature on effects of labour
turnover on firm performance. Existing research on the effects of worker replacement on
the performance of established firms is difficult to relate to young firms and shows ambigu-
ous results. Using structural identification to account for potential endogeneity of worker
replacement, this paper finds that worker replacement has negative effects on the productiv-
ity of young firms. It is disproportionally deleterious in high-tech manufacturing startups,
startups which employ highly educated employees, and in young firms above the median
size and age of the sample. A robustness check indicates that these results are driven by
firms with a high share of quits (by employees) on total separations. Replacements in the
sequel of dismissals, initiated by young firms themselves, seem to be less detrimental.
1
1 Introduction
Employment in young firms accounts for a considerable share of total employment in an econ-
omy. Besides jobs for the founders themselves, many startups start to create jobs for additional
employees early in their lives. By doing so, they compete against other young and established
firms for the most capable employees. However, while there is extensive research on the start-
up size and growth of young firms, only very little is known about labour turnover and potential
consequences of labour turnover for young firms. Labour turnover is a common phenomenon
in labour markets. On the aggregate level, as well as on the firm level, essentially more worker
flows than job flows can be observed (Hamermesh et al., 1996; Lane et al., 1996a). On the
aggregate level, this observation can be explained by a reallocation of workers between firms.
On the firm level, so called ”excess” labour turnover (also named churning by some authors)
occurs due to replacements of employees. When an employee is hired only to replace another
employee, no employment increase or decrease is triggered by the hiring. Labour turnover due
to worker replacement therefore occurs in excess of the labour turnover that would be necessary
due to a decline or an increase in employment size. Nevertheless, the replacement leads to two
worker flows - one into the firm and one out of the firm - that have to be managed. While the
inflow of human capital can bring new knowledge and ideas into a firm, the outflow of human
capital can be associated with a loss of important tacit knowledge and adjustment costs (e.g.
search costs for new employees or costs associated with the training of new employees). This
might have major implications for the future development of a young firm.
However, there are almost no economic studies that deal with the consequences of worker
replacement for the performance of young firms so far. Rather, studies reversely explain worker
replacement by firm performance and consider potential simultaneous effects, but do not unravel
these effects in more detail (Faberman and Nagypal, 2008). Studies on the consequences of
worker replacement on the performance of established firms are rare and their results are mixed.
While most authors conclude that negative effects predominate (Lane et al., 1996b; Burgess
et al., 2000a,b; Mamede and Mota, 2012), some papers report inverted u-shaped (Muller and
Peters, 2010) or even strictly positive relationships (Ilmakunnas et al., 2005). For several rea-
sons, the transferability of results about labour turnover in established firms is limited. Firstly,
small firms are not represented by most studies on labour turnover at all. To avoid frictions in
the analysis, most studies systematically exclude small firms from their samples and most young
firms are still small. Secondly, young firms might be more financially restricted in competing
with established firms for the best employees. Therefore, highly qualified employees with well
remunerated outside options might be responsible for a disproportionately high share of worker
quits in young firms. Thirdly, and maybe most importantly, hierarchies in new ventures can
be considered to be mostly flat and processes within those ventures to be often informal and
not yet standardised. One single employee might bear a much larger share of important tacit
knowledge than in larger established firms. The inflow or outflow of one single employee might
trigger much larger frictions for a new venture - compared to an established firm - therefore.
The focus of this paper is to unravel how worker replacement affects the productivity of
young firms. Due to the mixed findings about the consequences of worker replacement for
established firms the relationship is tested very generally at first. It is asked whether worker
replacement affects the productivity of young firms at all and if yes, whether an optimal re-
placement rate can be identified. Research on the extent of labour turnover in established firms
suggest differing results according to firm size, firm age, and industry (Lane et al., 1996b). This
2
seems reasonable since costs for on-the-job training of new employees should differ highly be-
tween different industries. For example, one might expect the replacement of employees to be
less costly in low-tech services firms compared to high-tech manufacturing firms. Firm age
and size might be associated with the availability of corporate knowledge about recruiting tech-
niques and the buildup of specialised human resource divisions to manage worker flows. This
could moderate the consequences of worker turnover. Thus, this paper analyses whether differ-
ent levels of excess turnover which can be expected over the industry, firm age and firm size
distributions reflect optimal behaviour or whether they bring forth consequences for firm per-
formance. The extent of labour turnover and its impact on firm performance are also likely to
depend on the capability of its managers to manage worker flows (Lane et al., 1996b) and on the
human capital of the employees involved in this flows (Mamede and Mota, 2012). Therefore,
this paper considers differing productivity effects dependent on the human capital of founders
and employees, i.e. the managerial and entrepreneurial experience of the founders and the qual-
ification structure of the workforce.
By addressing these questions, this paper contributes to two streams of literature: Firstly,
it is the first paper to study the effects of excess labour turnover on firm performance in the
specific context of young firms. Thus, it has important managerial implications for person-
nel strategies of young firms. Secondly, this paper helps to deepen the general understanding
of the effects of labour market flows on firm performance. Productivity is chosen as a very
broad and direct measure of firm performance. It should be less affected by mediating factors
than other potential performance measures like survival or profits. The likely endogeneity of
worker replacement is taken into account by using structural identification methods for produc-
tion functions proposed by Levinsohn and Petrin (2003) and Ackerberg et al. (2006). Possible
impacts of additional factors which might compromise identification, e.g. omitted variables in
the production function, are discussed and addressed in robustness checks. Using data from
the KfW/ZEW Start-Up Panel, a representative dataset on young German businesses, this paper
finds that excess worker turnover has negative effects on the productivity of young firms. A
robustness check indicates that these results are driven by firms with a high share of quits (by
employees) on total separations. Replacements in the sequel of dismissals, initiated by young
firms themselves, seem to be less detrimental. Worker replacement is disproportionally delete-
rious in high-tech manufacturing startups, startups which employ highly educated employees,
and in young firms above the median size and age of the sample.
2 Related literature
Two streams of literature offer especially interesting starting points for the empirical analysis
in this paper. Firstly, literature that studies the occurrence of labour turnover and identifies risk
factors for involuntarily high labour turnover rates. Secondly, streams of literature that deal
with the effects of (excess) labour turnover on firm performance.
Amongst other, match quality and firm performance can be risk factors for (involuntarily)
high labour turnover rates. Lane et al. (1996b) analyse the occurrence of labour turnover on the
firm level. They find that labour turnover rates vary substantially between firms of different size,
age, and industries but are persistent on the firm level. This persistence is explained by differing
abilities of managers to manage worker flows and to guarantee good matches between the firm
and the employee. With respect to this match quality, theoretical considerations of Jovanovic
3
(1979) show that the probability of separation is higher the shorter the respective job tenures
are. Jovanovic assumes the true quality of the match between employer and employee to be
revealed after a contract is made. Bad matches lead to fast separations, good matches survive
longer. Thus, the probability of separation is highest when job tenures are short and decreases
the longer tenures last. Other reasons for worker replacement due to bad matches might be little
experience in recruiting (Lane et al., 1996a) and strong employment growth in previous peri-
ods (Burgess et al., 2000a). In both cases the match quality between employer and employee
worsens because the management cannot assure the quality of each single match. This triggers
a need for subsequent replacement.
A second set of firm-level risk factors for involuntarily high labour turnover rates is re-
lated to the economic situation of the firm. Faberman and Nagypal (2008) present a theoretical
framework which relates firm productivity to worker flows. Their model predicts an increasing
outflow of human capital for firms with bad productivity. In their framework, firms are only
willing to replace quitting workers within certain productivity thresholds. However, while the
authors consider potential simultaneous effects between outflows of human capital and a further
deterioration of firm performance, they do not illustrate this aspect. Doing so might have impor-
tant implications for personnel strategies of firms. Related to this, in several empirical studies
low wages are shown to be important drivers of labour turnover. Amongst others, Martin (2003)
and Ilmakunnas and Maliranta (2005) demonstrate this relationship.
These risk factors seem to be relevant for young firms. Due to the young age, job tenures in
startups are short by definition. Often management experience of the founder(s) is still limited.
Corporate learning about hiring practises is still developing and specialised human resource de-
partments taking care of the selection of new employees might not be available. In addition,
employment growth, as well as potential downturns, are most pronounced at a very young age.
Finally, young firms’ possibilities on the labour market are limited. E.g. Brixy et al. (2007)
show that especially very young German startups often pay lower wages than established firms.
This might lead to an unfavourable outflow of human capital.
As a reflection of its uncertain consequences, the existing findings about the effects of excess
labour turnover on the performance of established firms are mixed. Muller and Peters (2010)
study the impact of worker replacement among R&D-personnel on firms’ innovative productiv-
ity. They find an inverted u-shaped relationship and conclude that moderate constant exchange
of R&D-personnel is favourable for the innovative capacity of established firms. Closest to
the research presented in this paper, Ilmakunnas et al. (2005) analyse the impact of worker
replacement on total factor productivity. They find a strictly positive impact of worker replace-
ment on productivity. Constantly replacing employees leads to productivity gains which the
authors attribute to an increase in match quality. Potential endogeneity of the worker replace-
ment measure, which might bias the results, is not addressed however. Burgess et al. (2000a)
analyse worker and job flows and find a complex dynamic interconnection between the two.
As discussed above, past job flows increase subsequent worker replacement. In turn, increas-
ing worker replacement effects subsequent growth negatively. Other empirical papers highlight
the interrelation between labour turnover and firm survival. Lane et al. (1996b) and in more
detail Burgess et al. (2000b), as well as Mamede and Mota (2012) analyse hazard rates and
report worker replacement to be associated with a lower probability of survival. According
to Lane et al. (1996b), these negative consequences of worker replacement on firm survival
are highly dependent on the industry sector. In addition, Burgess et al. (2000b) report nega-
4
tive consequences on survival especially for young firms with very high worker replacement
rates. Therefore, their study is one of the scarce exemptions to report insights on young firms.
However, the finding is only a side note in their study and it is not possible to assess if the re-
lationship is causal. In their recent working paper, Mamede and Mota (2012) confirm negative
effects of worker replacement on firm survival. Interestingly, they find the effects of worker
flows on firm survival to be more pronounced if highly skilled workers are involved. This holds
for both, negative effects due to separations of high-skilled employees and positive effects due
to high-skilled hirings.
3 Empirical setup
3.1 Model Choice
3.1.1 Productivity model
The empirical part of this study aims to establish whether worker replacement in young Ger-
man businesses affects their productivity. Therefore, productivity is modelled by an augmented
Cobb-Douglas production function and estimated in log-linear form:
lnYit = α + βlnLit + γlnKit + ρWRRit + θXit + uit
Y denotes real value added. L is the full-time equivalent size of the labour force, K denotes
capital, and WRR the worker replacement rate of a firm. X contains additional control variables:
the logarithmic age of a firm, a dummy variable indicating whether a firm is a limited liability
corporation, dummy variables to indicate whether the founder of a firm has prior experience as
an entrepreneur or has gained managerial experience as an employee1, per capita gross value
added in the federal state and the industry a firm operates in (to control for macroeconomic
influences on firm level productivity), as well as industry and time dummies. More details on
the measurement of variables are provided in Section 3.3 and Section 3.4.
3.1.2 Estimation & identification strategies
To identify causal relationships when estimating production functions, (potential) endogeneity
of variable production factors must be taken into account. Endogeneity can stem from simul-
taneity between a productivity shock and the use of variable inputs in the production function.
If a productivity shock is anticipated by the firm but unobserved or not measurable by the
economist, the estimated coefficients of variable inputs can be expected to be biased upwards
and those of fixed inputs to be biased downwards. This might happen since firms are assumed
to expand production in anticipation of a positive productivity shock and use more variable in-
puts in the production process (see Ackerberg et al. (2006) for a survey of procedures to tackle
the endogeneity problem). While classical production function literature is mainly concerned
with the endogeneity of the flexible production factor labour, the same reasoning holds for the
worker replacement measure in this study. In anticipation of a negative productivity shock not
observed in the productivity model, firms could decide to replace unproductive employees. Al-
ternatively, employees with high productivity could decide to voluntarily leave the firm, since
1For founding teams at least one of the founders must have managerial experience as an employee.
5
they are able to get better remunerated positions in more productive firms.2 Applying the same
reasoning for identification as used with labour input, the measure for worker replacement is
treated as an additional variable input into the production function.
Olley and Pakes (1996) suggest a semi-parametric procedure to identify the coefficients of
production functions. They propose to use structural assumptions about the optimal behaviour
of a firm in the production process and investments in the period the shock occurs as proxy vari-
able, to model the unobserved productivity shock in the estimation. However, this procedure is
only feasible if the proxy variable is non-zero and monotonically increasing in the productivity
shock. This makes the choice of the proxy variable problematic in the case of startups. Often,
young firms do not invest continuously but with gaps of several years in between. This leads
to many zero values (approximately 32 % of the observations in the sample used in this study)
in the investment series of a firm and makes the investment proxy hardly feasible. Levinsohn
and Petrin (2003) (LP henceforth) modify the procedure of Olley & Pakes and suggest to use
intermediate inputs (e.g. materials) as proxy. Since materials are needed to derive output as
value added in this study anyway and the proportion of zero values is distinctly smaller (ap-
proximately 2 % of the observations), the LP procedure seems to be a superior choice here.3
The basic intuitions and assumptions behind the LP procedure are the following (Levin-
sohn and Petrin (2003) and Petrin et al. (2004)): The capital stock is assumed to be a fixed
input in the current period t and not affected by a potential unobserved productivity shock. In
contrast, variable inputs into the production function (e.g. labour and intermediate inputs) are
considered to be freely adjustable in t and therefore potentially affected by an unobserved pro-
ductivity shock. The error term of the production function equation uit is assumed to consist
of two parts: uit = vit + ǫit. vit is known by the firm but unknown by the economist, ǫit is
an innovation unknown to both the firm and the economist. Assuming that a firm’s demand
for intermediate inputs depends on the fixed capital level and is strictly increasing in the pro-
ductivity shock, the demand for materials can be expressed as a function of capital and the
productivity shock: lnMit = f(vit, lnKit). Under the condition that the demand for materials
increases monotonically for all levels of the productivity shock this function can be inverted
to vit = f−1(lnMit, lnKit) to obtain a measure for the unobserved productivity shock. LP
suggest to approximate vit with a third order polynomial in lnMit and lnKit and to include
this proxy in the first stage of the estimation procedure to obtain unbiased OLS estimates of the
variable inputs.4 Assuming that the productivity shock follows a first-order Markov process, the
coefficients for capital and materials can be disentangled in the second step of the estimation
procedure using a non-linear least squared regression. Finally, the standard errors are derived
using a bootstrap procedure to account for estimations on both stages.
The most crucial assumption might be that a firm’s labour input is seen as freely variable.
In general, this might be hard to justify in the rigid German labour market. However, there is
an exclusion from dismissal protection for small firms in German law, known as ”small firm
2Such behaviour is explained within the theoretical framework of Faberman and Nagypal (2008)3Ackerberg et al. (2006) discuss the limitations of the approaches suggested by Olley and Pakes (1996) and
Levinsohn and Petrin (2003) and suggest a modified estimation method. Their procedure is implemented as a
robustness check.4A (smoothed) graph of the relationship between materials, capital and the estimated productivity shock is
shown in Figure 1 in the appendix. The estimated productivity shock is measured on the vertical axis, capital and
materials on the horizontal axes. The use of materials is higher for all levels of capital input and all levels of the
estimated productivity shock.
6
clause”. Firms which employ ten employees or less are not restricted by dismissal protection in
Germany. Roughly 90 % of the sample used in this study fall under this special legislation.5. To
address this problem more formally, results are checked for robustness using an estimation ap-
proach suggested by Ackerberg et al. (2006). This approach allows lagged variable inputs into
the production function to influence productivity and relaxes the timing assumptions therefore
Parrotta and Pozzoli (2012). This might also offer a more realistic scenario for the impact of
worker replacement on productivity therefore. In the following, presented baseline estimations
are conducted by pooled OLS with cluster robust standard errors and than confirmed by the
semi-parametric approach suggested by LP. All estimations are done in Stata. Results for LP
approach are estimated using the routine implemented by Petrin et al. (2004). Results do not
change qualitatively when models are re-estimated by the procedure suggested by Ackerberg
et al. (2006).6
An important limitation of this study is the inability to control for human resource manage-
ment practises implemented by the firms or the existence of an own human resource depart-
ment. These factors could influence both the degree of labour turnover and the productivity of
the employees (e.g. by influencing motivation) and impair the identification of a causal effect of
worker turnover on productivity. However, it is argued that such factors are covered to a large
extent by controlling for the skill level and professional experience of the founders, firm age
and firm size.
3.2 Data
The empirical analysis is based on the first six survey waves of the KfW/ZEW Start-Up Panel7,
a yearly survey of newly established firms representative for Germany. The panel was started
in 2008 with a survey of about 5,500 firms founded as legally independent companies between
2005 and 2007. A follow-up survey of firms which have already participated in the survey is
conducted in each of the subsequent years since. In addition, each year the dataset is enhanced
with a sample of new firms which have been founded within the last three years. By now, the
database contains information on 15,300 firms founded between 2005 and 2012. The sample is
a stratified random samples drawn from the population of all firm creations which are recorded
by Creditreform (Germany’s largest credit rating agency). The stratification criteria are the year
of foundation, the industry and funding by the KfW. The main goal of the stratification is an
oversampling of high-tech startups which allows to conduct separate analyses for this groups
of startups.8 Detailed information about the founders, their human capital, the firms’ labour
demands, and indicators of firm performance are retrieved by computer-assisted telephone in-
terviews. More detailed information on the survey design of the dataset is provided in Fryges
et al. (2010).
There are some restrictions on the data used for estimation in the empirical section of this
paper. Since costs of materials can only be identified from the third survey wave onwards, only
data points from 2009 to 2012 are used. Since calculating the worker replacement measure as
a rate involves dividing by the average number of employees, all estimations can only be done
5As a further robustness check, firms that do not fall under the ”small firm clause” were deleted from the
sample. This does not alter the results.6Results of the robustness check are available from the author upon request.7The KfW/ZEW Start-Up Panel is a joint project of the Centre for European Economic Research (ZEW), the
KfW-Bankengruppe, a publicly-owned bank, and Creditreform (Germany’s largest credit rating agency)8Stratification criteria are controlled for in all regressions and extrapolated descriptive statistics.
7
for startups with at least one employee. All variables measured in monetary units are converted
to 2010 prices using a GDP price deflator series provided by the German Federal Statistical
Office. Additional data, for a gross value added index series, is taken from data provided by the
Statistical Office of the Federal State of Baden-Wurttemberg. To adjust for erroneous data, the
largest percentile of all variables measured in monetary values and of the worker replacement
rate is cut.
3.3 Worker replacement measure
To measure worker replacement, a slightly adjusted version of the churning rate introduced
by Burgess et al. (2000a) is used. Burgess et al. consider all worker flows in and out of a
firm that occur in a period and define churning flows as the difference between worker flows
and job reallocation. Worker flows equal all hirings plus all separations. They measure the
total amount of worker turnover that occurs in a firm in a period. The job reallocation in
a firm is measured as the absolute value of the difference between hirings and separations
(Job reallocation = |Hirings − Separations|). Thus, it measures the net change of a firm’s
employment size. Job reallocation equals the lower bound of worker flows that are necessary to
adjust the labour force of a firm for changes in it’s labour demand. All worker flows that exceed
this lower bound occur just to replace existing workers. May it be because workers quit, because
they are dismissed or because they leave the workforce (e.g. because of retirement). Therefore,
churning is employment-neutral and allows to separate labour turnover, due to an adjustment to
economic conditions, and additional ”exceeding” labour turnover. The employment neutrality
makes the churning measure particularly useful for the analysis of young firms, which are often
subject to enormous frictions in their development. Since their need for a steady adjustment
of labour inputs might be very high, analysing firm performance based on turnover measures
incorporating all worker flows might less beneficial.
The aim of Burgess et al. (2000a) is to separate all worker flows that occur in a firm (in
one period) into categories. The replacement of one worker leads to two churning flows: One
hiring and one separation. The focus of this study is on the effects of worker replacement on
firm performance however. Therefore, the churning rate is divided by two, to get a measure for
the proportion of replaced employees, which is straightforward to interpret on the firm level.9
Following Burgess et al. (2000a), to relate worker flows to the size of a firm, rates are calculated
by dividing total churning flows by average employment in a period.
Worker Replacement Rate it =Hiringsit + Separationsit − |Hiringsit − Separationsit|
2 ∗ 0.5 ∗ (#Employeesit +#Employeesit−1)
3.4 Descriptive statistics
Descriptive statistics for variables that enter the estimations are reported in Table 1. Variables
measured in monetary units enter the estimations in logarithmic form but are reported as abso-
lute (deflated) values in the descriptive statistics table.
On average, the startups in the regression sample generate a real value added of 350,000
EUR per year. A histogram of the dependent variable, logarithmic real value added, is provided
9A comparable measure is used e.g. by Albaek and Sorensen (1998) and Muller and Peters (2010).
8
in Figure 2 in the appendix. The distribution of logarithmic real value added seems to be reason-
ably close to the (indicated) normal distribution. Using OLS as baseline model and at the first
stage of the semiparametric estimator seems a justifiable choice with respect to the distribution
of the dependent variable therefore.
The average size of the capital stock is 156,000 EUR. Following Levinsohn and Petrin
(2003), the capital stock for a period is calculated as the sum of the depreciated capital stock of
the last period and investments in the current period. Since no detailed information on deprecia-
tion rates is available for the firms, the capital stock of the forgone period is always depreciated
by 10 %.10. The startups engage on average 5.6 employees (full-time equivalent workforce size
including founders). The employment size is widespread and varies between 1 and 143 em-
ployees. Since founders contribute to value added they are included in the measure for labour
input. However, they are excluded in the calculation of the worker replacement measure which
is derived for dependently employed workers only and, in accordance with most literature on
labour turnover, measured in headcounts. 27 % of the observations in the sample exhibit pos-
itive worker replacement rates. The average worker replacement rate is 11 %. More detailed
information on the extent and the distribution of worker replacement in German startups is pro-
vided in the appendix.11 The second and third sections of Table 1 demonstrate that firms with
positive WRR are larger in terms of employment size and value added than firms with no worker
replacement. There do not seem to be distinct differences in firm age however.
Firms are on average 3.7 years old (at the end of the respective reporting period). 46 % of
the observations are from limited liability companies. 47 % of the observations are from firms
whose founders are managerially experienced. Managerial experience is measured by a dummy
variable which indicates whether the founder (or at least one of the founders in the team) worked
in an executive position in dependent employment directly prior to the foundation of the startup.
37 % of the founders have entrepreneurial experience, which means that the founder (or at least
one founder in the team) was self-employed before. On average 22 % of the workforce of the
startups are highly qualified, i.e. they hold a university degree.
To control for macroeconomic conditions, real gross value added (GVA) in the federal state
a firm operates in is included as an index series. To consider industry and period specific differ-
ences in the macroeconomic conditions a firm operates under, GVA is differentiated by year and
a rough three sector industry classification (i.e. manufacturing, services & retail and construc-
tion). In addition, industry and year dummies are included. While data points are distributed
relatively evenly over the four years from 2009 to 2012, the distribution of firms over different
industries reflects the stratification criterion reported in Section 3.2. High-tech firms (first four
industries) account for roughly 40 % of the estimation sample. About 14 % of the data points
are from high-tech manufacturing firms (first two industries), another 27 % of the data points
are from high-tech services firms (third and fourth industries).
10Other depreciation rates between 10 % and 15 % have been tested as a robustness check and do not alter the
results remarkably.11Please note that the deviation in mean WRR, compared to the extrapolated values reported in the appendix, is
a consequence of the overrepresentation of high-tech startups in the sample. On average, high-tech startups report
lower worker replacement rates.
9
Table 1: Descriptive Statistics
Complete regression sample No worker replacement Positive worker replacement
Variables Obs. Mean Std. Dev. Min. Max. Obs Mean Std. Dev. Obs Mean Std. Dev.
Real value added 5613 350262.30 427760 987.77 4673189 4124 313616.60 392824.10 1489 451757.60 498503.10
Labour (# employees) 5613 5.63 6.23 1 143 4124 4.64 4.05 1489 8.37 9.54
Capital 5613 156310.30 294328.80 810 11000000 4124 142426.40 291976.30 1489 194764.00 297481.80
Materials 5613 186087.90 351711 101.03 5257337 4124 165469.60 310286.30 1489 243193.40 441958.00
Churning y/n 5613 0.27 0.44 0 1 4124 0.00 0.00 1489 1.00 0.00
Churning rate 5613 0.11 0.25 0 2 4124 0.00 0.00 1489 0.40 0.34
Age of firm 5613 3.72 1.64 1.08 8 4124 3.78 1.65 1489 3.57 1.60
Limited liability corporation 5613 0.46 0.50 0 1 4124 0.45 0.50 1489 0.51 0.50
Managerial experience 5613 0.47 0.50 0 1 4124 0.48 0.50 1489 0.46 0.50
Entrepreneurial exp. 5613 0.37 0.48 0 1 4124 0.36 0.48 1489 0.41 0.49
Highly qual. workers 5613 0.22 0.33 0 1 4124 0.23 0.34 1489 0.20 0.31
GVA p.c. (industry/state) 5613 102.65 7.38 78.50 129.32 4124 102.70 7.42 1489 102.51 7.27
Cutting-edge tech. manuf. 5613 0.08 0.27 0 1 4124 0.08 0.28 1489 0.06 0.23
High-tech manufacturing 5613 0.06 0.24 0 1 4124 0.06 0.24 1489 0.06 0.23
Technology-intensive services 5613 0.20 0.40 0 1 4124 0.22 0.41 1489 0.15 0.35
Software supply and consult. 5613 0.07 0.25 0 1 4124 0.07 0.25 1489 0.06 0.24
Non-high-tech manufacturing 5613 0.13 0.34 0 1 4124 0.13 0.34 1489 0.14 0.35
Skill-intensive services 5613 0.07 0.25 0 1 4124 0.07 0.25 1489 0.06 0.24
Other business-oriented serv. 5613 0.05 0.22 0 1 4124 0.04 0.21 1489 0.08 0.27
Consumer-oriented services 5613 0.11 0.31 0 1 4124 0.09 0.29 1489 0.17 0.37
Construction 5613 0.11 0.31 0 1 4124 0.10 0.30 1489 0.12 0.32
Wholesale and retail trade 5613 0.13 0.33 0 1 4124 0.13 0.34 1489 0.12 0.32
Year 2009 5613 0.20 0.40 0 1 4124 0.20 0.40 1489 0.19 0.40
Year 2010 5613 0.24 0.42 0 1 4124 0.23 0.42 1489 0.24 0.43
Year 2011 5613 0.27 0.44 0 1 4124 0.27 0.44 1489 0.28 0.45
Year 2012 5613 0.29 0.46 0 1 4124 0.30 0.46 1489 0.29 0.45
10
4 Empirical results
General results Main regression results are shown in Table 2. At first, the baseline model,
without the measure for worker replacement, is estimated by pooled OLS with cluster robust
standard errors (first column). Coefficients for labour and capital input are in a plausible range.
Older startups, limited liability corporations, startups whose founders have prior management
experience as employees, and startups which operate in a better macroeconomic environment
(approximated by gross value added in the federal state and the industry) have significantly
higher productivity. When the baseline model is re-estimated with the structural LP approach
(second column), the labour coefficient decreases clearly and the capital coefficient increases
slightly. This is the expected effect when endogeneity of the variable inputs is controlled for by
the LP procedure and indicates the functioning of the approach in the setup of this study.
To address research question if worker replacement affects productivity, worker replace-
ment is integrated in the model as an additional variable (third column for pooled OLS results,
fourth column for LP). The estimated semi-elasticities indicate that worker replacement has a
significant negative effect on the productivity of young firms. All other coefficients of the model
remain stable. As expected, the coefficient of WRR increases slightly (from -0,106 to -0.095)
when the model is estimated by the LP procedure but the qualitative implications do not change
once potential endogeneity is taken into account. Since differences between pooled OLS and
LP estimates are only minor for all specifications, only the slightly more conservative LP results
are discussed for all remaining regressions. Corresponding OLS results are provided in Tables
5 and 6 in the appendix. To check for turning points in the relationship between labour turnover
and productivity, squared WRR is considered in addition (fifth column). The coefficient of the
squared term is insignificant, thus the negative relationship seems to hold over the whole distri-
bution of the WRR. This result can be confirmed when worker replacement is integrated only as
a dummy variable (sixth column). In general, firms with a positive WRR exhibit significantly
lower productivity.
Sensitivity analysis A more detailed sensitivity analysis is conducted to better explain the
origin of the observed negative relationship between worker replacement and productivity (see
Table 3). The sensitivity analysis is performed by split sample analyses to asses the influence of
subgroups of the sample on regression results. In addition, versions of the model are extended
by interaction terms between WRR and several measures that potentially mediate the effect of
labour turnover on productivity. As a first step of the sensitivity analysis, the regression sample
is split into the larger and the smaller half of young firms according to their employment size
(first and second column). This is done for two reasons. Firstly, to check for size effects in
the relationship between labour replacement and productivity in general. Secondly, to make
sure the results are not driven by basis effects in the calculation of the worker replacement rate.
Since some of the firms are very small, the replacement of a single employee can trigger large
WRR. The results rule out pollution through basis effects. The negative relationship between
worker replacement and productivity seems to be driven mainly by firms above the median size
of the sample.
11
Table 2: Estimated Productivity Effects of Worker Replacement - Main Results
Dependent variable: Baseline - OLS Baseline - LP With WRR - OLS With WRR - LP WRR & WRR-sq. - LP WRR y/n - LP
Real value added Coef. S.E. Coef. S.E. Coef. S.E. Coef. S.E. Coef. S.E. Coef. S.E.
Labour (# employees) 0.772 (0.021)*** 0.702 (0.020)*** 0.774 (0.021)*** 0.705 (0.020)*** 0.708 (0.021)*** 0.712 (0.021)***
Capital 0.231 (0.015)*** 0.237 (0.082)*** 0.231 (0.015)*** 0.238 (0.082)*** 0.237 (0.081)*** 0.237 (0.082)***
Worker replacement rate -0.106 (0.047)** -0.095 (0.045)** -0.211 (0.097)**
Worker replacement sq. 0.107 (0.094)
Worker replacement y/n -0.048 (0.024)**
Age of firm 0.117 (0.027)*** 0.090 (0.028)*** 0.112 (0.027)*** 0.085 (0.028)*** 0.085 (0.028)*** 0.087 (0.028)***
Limited liability corporation 0.256 (0.031)*** 0.189 (0.030)*** 0.254 (0.031)*** 0.188 (0.030)*** 0.188 (0.030)*** 0.188 (0.030)***
Managerial experience 0.129 (0.026)*** 0.117 (0.025)*** 0.127 (0.026)*** 0.116 (0.025)*** 0.114 (0.025)*** 0.116 (0.025)***
Entrepreneurial experience 0.010 (0.029) 0.011 (0.027) 0.012 (0.029) 0.013 (0.027) 0.012 (0.027) 0.012 (0.027)
Highly qual. workers (share) 0.063 (0.045) 0.088 (0.041)** 0.063 (0.045) 0.088 (0.041)** 0.087 (0.041)** 0.087 (0.041)**
GVA p.c. (industry/state) 0.003 (0.002)* 0.002 (0.002) 0.003 (0.002)* 0.002 (0.002) 0.003 (0.002) 0.002 (0.002)
Cut.-edge tech. manuf. -0.116 (0.056)** 0.019 (0.058) -0.120 (0.056)** 0.015 (0.057) 0.014 (0.057) 0.015 (0.057)
High-technology manuf. -0.194 (0.068)*** -0.129 (0.061)** -0.197 (0.068)*** -0.132 (0.061)** -0.133 (0.061)** -0.133 (0.061)**
Tech.-intensive services 0.036 (0.053) 0.193 (0.053)*** 0.032 (0.053) 0.189 (0.053)*** 0.188 (0.053)*** 0.190 (0.053)***
Software supply and cons. -0.067 (0.065) 0.167 (0.065)** -0.069 (0.065) 0.164 (0.065)** 0.162 (0.065)** 0.165 (0.065)**
Non-high-tech manuf. -0.157 (0.052)*** -0.053 (0.050) -0.160 (0.052)*** -0.056 (0.050) -0.056 (0.050) -0.054 (0.050)
Skill-intensive services -0.068 (0.059) 0.134 (0.063)** -0.069 (0.059) 0.132 (0.062)** 0.133 (0.062)** 0.134 (0.063)**
Other business-or. serv. -0.107 (0.070) 0.078 (0.071) -0.100 (0.070) 0.083 (0.072) 0.083 (0.072) 0.082 (0.072)
Consumer-or. services -0.221 (0.059)*** -0.037 (0.058) -0.215 (0.059)*** -0.032 (0.058) -0.030 (0.058) -0.031 (0.058)
Construction -0.076 (0.053) -0.024 (0.050) -0.074 (0.053) -0.022 (0.050) -0.022 (0.050) -0.024 (0.050)
Year 2010 -0.029 (0.029) -0.015 (0.027) -0.030 (0.029) -0.016 (0.027) -0.015 (0.027) -0.015 (0.027)
Year 2011 -0.005 (0.032) -0.009 (0.031) -0.005 (0.032) -0.009 (0.031) -0.009 (0.031) -0.008 (0.031)
Year 2012 0.003 (0.032) 0.010 (0.032) 0.002 (0.032) 0.010 (0.032) 0.010 (0.032) 0.011 (0.032)
Constant Yes Yes Yes Yes Yes Yes
Observations / R-squared 5,613 / 0.514 5,613 5,613 / 0.514 5,613 5,613 5,613Notes: *** 1%, ** 5 %, * 10 %. Cluster robust / bootstrapped standard errors in parentheses. Additional control variable in all regressions: Funding by the KfW bank.
12
Table 3: Estimated Productivity Effects of Worker Replacement - Sensitivity Analysis 1
Dependent variable: Empl. > MED - LP Empl. <= MED - LP Industries - LP Firm Age > MED - LP Firm Age <= MED - LP Firm Age det. - LP
Real value added Coef. S.E. Coef. S.E. Coef. S.E. Coef. S.E. Coef. S.E. Coef. S.E.
Labour (# employees) 0.724 (0.026)*** 0.503 (0.048)*** 0.705 (0.020)*** 0.686 (0.030)*** 0.713 (0.024)*** 0.706 (0.020)***
Capital 0.326 (0.111)*** 0.376 (0.114)*** 0.237 (0.081)*** 0.117 (0.096) 0.385 (0.140)*** 0.234 (0.079)***
Worker replacement rate -0.196 (0.054)*** -0.052 (0.061) -0.077 (0.062) -0.165 (0.069)** -0.049 (0.058) -0.073 (0.068)
WRR # HTM -0.153 (0.117)
WRR # HTS -0.013 (0.137)
WRR # LTM 0.056 (0.178)
WRR # Constr -0.048 (0.112)
WRR # Age 3 -0.016 (0.092)
WRR # Age 4 0.068 (0.101)
WRR # Age 5 -0.035 (0.148)
WRR # Age 6 -0.199 (0.122)
WRR # Age 7 -0.344 (0.187)*
WRR # Age 8 -0.047 (0.144)
Age of firm 0.063 (0.027)** 0.076 (0.041)* 0.085 (0.028)*** -0.114 (0.082) 0.190 (0.044)*** 0.095 (0.030)***
Limited liability corporation 0.193 (0.032)*** 0.193 (0.047)*** 0.188 (0.030)*** 0.165 (0.036)*** 0.200 (0.043)*** 0.187 (0.030)***
Managerial experience 0.050 (0.030)* 0.168 (0.038)*** 0.115 (0.025)*** 0.123 (0.030)*** 0.111 (0.035)*** 0.116 (0.025)***
Restarter 0.011 (0.026) 0.005 (0.045) 0.012 (0.027) 0.050 (0.031) -0.015 (0.037) 0.012 (0.027)
Highly qual. workers (share) 0.122 (0.056)** 0.089 (0.064) 0.089 (0.041)** 0.066 (0.053) 0.109 (0.061)* 0.088 (0.041)**
Constant / All Controls Yes Yes Yes Yes Yes Yes
Observations / R-squared 2,787 2,826 5,613 2,737 2,876 5,613Notes: *** 1%, ** 5 %, * 10 %. Bootstrapped standard errors in parentheses. Additional control variable in all regressions: Funding by the KfW bank.
13
A similar pattern arises when the sample is split into the older and the younger half of firms
to control for the influence of firm age on the relationship between excess turnover and produc-
tivity. Results are only significant for firms of above median age. Clearly, the overlap between
the older and the larger half of the sample can be expected to be high. To control for the age
effect in more detail, the WRR is interacted with the firm age categorised into years (sixth col-
umn). The sum of the main and the interaction effect of worker replacement on productivity is
only significant for six- and seven-year-old firms (6 years: 1 % level, 7 years: 5% level). Since
the group of eight-year-old startups in the regression sample is very small this could blur effects
at the upper end of the firm age distribution. So, despite the smooth evolution of the WRR over
the first period of the life-cycle (see Section 2), excess labour turnover is most deleterious when
firms grow.
The extent of labour turnover differs distinctly between firms of different industries. To
address the question whether the effect of labour turnover on productivity differs between in-
dustries, the WRR is interacted with five sector dummies (third column). The sum of effects
is only significant for high-tech manufacturing startups (5 % level). So despite the descriptive
result that high-tech manufacturing startups have the lowest worker replacement rates of all
sectors, worker replacement seems to be most harmful to them. While turnover rates in other
sectors are more pronounced, their effects on productivity are not significant. However, the sum
of the direct effect and the interaction effect is in no case positive.
As a further sensitivity check it is analysed how founders and employees human capital
moderate the effect of excess labour turnover on productivity (see Table 4). As discussed in
Section 2, the experience of a firm’s managerial staff can assure good quality of the employer-
employee matches and thus, limit involuntary labour turnover. The mediating effects of rele-
vant prior experience of the founder(s) is approximated in two ways. Firstly, by interacting the
WRR with the dummy variable indicating whether the founder (or one founder in the team)
has prior managerial experience in dependent work (first column). Secondly, by interacting the
WRR with the dummy variable indicating if the founder (or one founder in the team) has prior
entrepreneurial experience (second column). The extension of the productivity model by the in-
teraction between WRR and managerial experience does not provide any insights. In contrast,
results indicate that mainly startups whose founders have no experience as entrepreneurs are re-
sponsible for the significant negative overall effect of worker replacement on productivity (the
sum of the direct effect and the interaction effect is not significantly different from zero). This
result is reinforced if prior entrepreneurial experience is split into two categories, successful
prior entrepreneurs and failed prior entrepreneurs.12 In both cases the sum of the direct effect
of WRR and the interaction effect is not significantly different from zero. However, the positive
interaction effect is clearly more distinct for successful entrepreneurs.
Finally, since Mamede and Mota (2012) find an amplifying effect of worker flows involving
highly skilled workers on firm performance, it is tested whether this holds for the effect of
worker replacement on the productivity of young firms as well. Unfortunately, the data used
in this study does not allow to observe the qualification of a replaced worker directly. To shed
some light on the question if labour turnover is more deleterious for young firms if highly skilled
labour is involved, an interaction term between the WRR and a dummy variable indicating
12Entrepreneurs are considered successful if they continued their old business, passed it on to friends or fam-
ily members who continued it, or if they sold it profitably. Entrepreneurs are considered unsuccessful (failed
entrepreneurs) if their prior business was divested or went bankrupt.
14
Table 4: Estimated Productivity Effects of Worker Replacement - Sensitivity Analysis 2
Dependent variable: Manag. Exp. - LP Entrepr. Exp. - LP Entrepr. Exp. (+-) - LP Qual. Empl. - LP
Real value added Coef. S.E. Coef. S.E. Coef. S.E. Coef. S.E.
Labour (# employees) 0.705 (0.020)*** 0.705 (0.020)*** 0.705 (0.018)*** 0.706 (0.021)***
Capital 0.238 (0.081)*** 0.237 (0.082)*** 0.240 (0.084)*** 0.236 (0.081)***
Worker replacement rate -0.095 (0.063) -0.127 (0.062)** -0.127 (0.067)* -0.061 (0.045)
WRR # Managerial experience -0.000 (0.087)
WRR # Entrepreneurial exp. 0.074 (0.086)
WRR # Successful entrepreneur 0.096 (0.108)
WRR # Failed entrepreneur 0.008 (0.122)
WRR # Hi. qual. work (y/n) -0.093 (0.098)
Age of firm 0.085 (0.028)*** 0.085 (0.028)*** 0.085 (0.028)*** 0.086 (0.028)***
Limited liability corporation 0.188 (0.030)*** 0.188 (0.030)*** 0.189 (0.031)*** 0.189 (0.031)***
Managerial experience 0.116 (0.027)*** 0.116 (0.025)*** 0.121 (0.025)*** 0.115 (0.025)***
Entrepreneurial experience 0.013 (0.027) 0.004 (0.028) 0.012 (0.027)
Successful entrepreneur 0.009 (0.035)
Failed entrepreneur 0.006 (0.035)
Highly qual. workers (share) 0.088 (0.041)** 0.088 (0.041)** 0.089 (0.042)** 0.098 (0.041)**
Constant / All Controls Yes Yes Yes Yes
Observations / R-squared 5,613 5,613 5,576 5,613
Notes: *** 1%, ** 5 %, * 10 %. Bootstrapped standard errors in parentheses. Additional control variable in all regressions: Funding by the KfW bank.
whether a firm employs highly qualified personnel is included (fourth column). The results are
in line with the existing findings for established firms. The sum of the direct effect of WRR
and the interaction effect is significant (on a 10 % level) while the direct effect is not significant
alone. Excess labour turnover is most deleterious in young firms that employ highly qualified
personnel. This supports the suspicion that highly qualified personnel, with well remunerated
outside options in other firms, might be responsible for a important share of quits in young
firms.
Quits vs. dismissals Two possible explanations are in line with the observed negative effect
of worker replacement on productivity. On the one hand, the negative productivity effect could
stem from the replacement of employees who were dismissed by the firm. The negative effect
would either reflect an underestimation of the adjustment costs associated with replacements
or a bad quality of the original match then. On the other hand, the negative effect might stem
from quits of employees. It seems reasonable that the most productive employees might be the
first to leave a firm voluntarily and that young firms might find it difficult to replace this out-
flow equivalently. Both explanations seem to be convincing in the context of young firms, with
comparably low power on the labour market and limited experience in recruiting.
To shed light on possibly different impacts of quits and dismissals a robustness check is
conducted. Unfortunately, the data available neither allows to distinguish if a separation was a
quit or a dismissal on the individual level, nor to identify which employee was replaced in the
case of multiple separations in one period.13 However, for the years 2007 and 2012 information
on the total number of separations due to quits and due to dismissals is available for the firms
13e.g. if there are two separations and one hiring, it cannot be distinguished which one of the two separations
was replaced and which one not.
15
which took part in the corresponding waves. To not loose to many observations for the robust-
ness check, the share of quits on all separations is treated constant over time on the firm level.
In case the information is available for a firm in 2007 and 2012, the share of quits is set to the
closest observed value available for each year. Table 7 (OLS results) and Table 8 (correspond-
ing LP results) in the appendix show the estimation results for the robustness check. Due to
the incomplete availability of data on quits and dismissals and the strong assumption about the
constancy of the share of quits over time, the results of the robustness check should be treated
cautiously.
In a first step (first columns of Table 7 and Table 8 respectively) the model is re-estimated
with the smaller set of observations that the share of quits is available for. Using both procedures
most estimated coefficients remain in the same range as in the model including the full set of
observations. OLS results seem to be more stable over the different specifications of the robust-
ness check. Since they do not lead to qualitatively different conclusions than the LP procedure,
the remaining discussion is based on the OLS results. In a second step the sample is split into
observations for which dismissals account for 50 % of the separations or more (second row)
and observations with more than 50 % quits (third row). The results indicate that the negative
effects of worker replacement on productivity only hold for firms with high shares of quits, not
for firms with mainly firm initiated separations. This result can be confirmed if the information
on the share of quits is introduced in the model by interaction terms (fourth and fifth row). Only
the sum of the direct and the interaction effect is (negatively) significant (on a 1 % level). This
result holds independently of interacting worker replacement with a dummy, indicating a share
of quits on all separations of more than 50 % (fourth row), or the actual share of quits (fifth row).
5 Conclusion
The analyses in this paper show that the overall effect of worker replacement on the productivity
of young firms is negative. This result is independent of the extent of observed worker replace-
ment. In a more detailed sensitivity analysis, high-risk groups for harmful worker replacement
are identified: young firms of above median employment size and median age of the sample,
young firms that employ highly educated employees, manufacturing startups in high-tech sec-
tors and startups whose founders do not have any prior experience as entrepreneurs. Conditions
under which worker replacement triggers positive effects on productivity cannot be identified.
A robustness check which has to rely on a restricted set of data indicates that the negative pro-
ductivity effects stem from replacements of employees who quit their jobs. For firms which
initiate most of the separations themselves, negative effects are not measurable. However, even
in the case of dismissals, a positive impact of worker replacement on firm performance cannot
be found.
A one unit increase in worker replacement is associated with a decrease in value added of
approximately 10 %. This results differs only minimally dependent on the regression method
used. Expressed in mean terms, the replacement of a young firms entire workforce during one
period leads to a value added loss of approximately 35,000 EUR. A more graspable example
might be the following: if the median employment size young firm in the sample, with four
full-time equivalent employees, replaces (or has to replace) two of its four employees during
a period, this is associated with an average loss in value added of approximately 17,500 EUR.
These losses become up to three times as high in firms where most separations are initiated by
16
quits of employees.
The negative effects of worker replacement on young firms’ performance are in line with
the theoretical considerations and might be explained best by a comparably low market power
of young firms on the labour market. It seems to hold true that, in many cases, well educated
employees leave young firms for more attractive other options. This does most harm firms in
high-tech manufacturing sectors which depend crucially on a highly skilled workforce. Interest-
ingly, the negative effect is strongest as firms grow. With the increasing size it seems to becomes
more difficult for the founders to manage the human resources of their startup adequately. This
is underpinned by the finding that founders with prior entrepreneurial experience succeed in
avoiding harmful worker replacement.
From these results important managerial guidelines can be derived. The managerial staff of
high-tech manufacturing start-ups should consider to increase the effort put into avoiding excess
labour turnover among highly skilled employees. A useful way to go might be to the develop
or reconsider their firms’ strategies to increase the job embeddedness of highly educated em-
ployees. Especially for founders without prior entrepreneurial experience, it might be valuable
to consider seeking the assistance of personnel consultants to develop better human resource
management strategies.
References
Daniel Ackerberg, Kevin Caves, and Garth Frazer. Structural identification of production func-
tions. MPRA Paper 38349, University Library of Munich, Germany, December 2006.
Karsten Albaek and Bent E. Sorensen. Worker flows and job flows in danish manufacturing
1980-1991. Economic Journal, 108,451, Nov.1998:1750–1771, 1998.
Udo Brixy, Susanne Kohaut, and Claus Schnabel. Do newly founded firms pay lower wages?
first evidence from germany. Small Business Economics, 29(1/2):161 – 171, 2007.
Simon Burgess, Julia Lane, and David Stevens. Job flows, worker flows, and churning. Journal
of Labor Economics, 18(3):473 – 502, 2000a.
Simon Burgess, Julia Lane, and David Stevens. The reallocation of labour and the lifecycle of
firms. Oxford Bulletin of Economics & Statistics, 62(5):885 – 907, 2000b.
Jason Faberman and va Nagypal. Quits, worker recruitment, and firm growth: theory and
evidence. Working Papers 08-13, Federal Reserve Bank of Philadelphia, 2008.
Helmut Fryges, Sandra Gottschalk, and Karsten Kohn. The kfw/zew start-up panel: Design
and research potential. Schmollers Jahrbuch/Journal of Applied Social Sciences Studies,
European Data Watch 130:117–131, 2010.
John Haltiwanger, Henry Hyatt, Erika McEntarfer, and Liliana Sousa. Job creation, worker
churning, and wages at young businesses. Ewing Marion Kauffman Foundation, Nov 2012.
Daniel S. Hamermesh, Wolter H. J. Hassink, and Jan van Ours. Job turnover and labor turnover:
A taxinomy of employment dynamics. Annales d’Economie et de Statistique, (41-42):21–40,
1996.
17
Pekka Ilmakunnas and Mika Maliranta. Worker inflow, outflow, and churning. Applied Eco-
nomics, 37(10):1115 – 1133, 2005.
Pekka Ilmakunnas, Mika Maliranta, and Jari Vainiomki. Worker turnover and productivity
growth. Applied Economics Letters, 12(7):395 – 398, 2005.
Boyan Jovanovic. Job matching and the theory of turnover. Journal of Political Economy, 87
(5):972 – 990, 1979.
Julia Lane, David Stevens, and Simon Burgess. Worker and job flows. Economics Letters, 51
(1):109 – 113, 1996a.
Julia I. Lane, Alan G. Isaac, and David W. Stevens. Firm heterogeneity and worker turnover.
Review of Industrial Organization, 11:275–291, 1996b.
James Levinsohn and Amil Petrin. Estimating production functions using inputs to control for
unobservables. Review of Economic Studies, 70(2):317 – 341, 04 2003.
Ricardo Mamede and Daniel Mota. Labour mobility and firm survival in knowledge intensive
services. Dinamia’cet Working Paper 2012/08, 2012.
Christopher Martin. Explaining labour turnover: Empirical evidence from uk establishments.
LABOUR, 17(3):391 – 412, 09 2003.
Kathrin Muller and Bettina Peters. Churning of r&d personnel and innovation. ZEW Discussion
Papers 10-032, ZEW - Zentrum fr Europische Wirtschaftsforschung / Center for European
Economic Research, 2010.
G. Steven Olley and Ariel Pakes. The dynamics of productivity in the telecommunications
equipment industry. Econometrica, 64:1263 – 1297, 1996.
Pierpaolo Parrotta and Dario Pozzoli. The effect of learning by hiring on productivity. The
RAND Journal of Economics, 43(1):167–185, 2012.
Amil Petrin, Brian P. Poi, and James Levinsohn. Production function estimation in stata using
inputs to control for unobservables. Stata Journal, 4(2):113 – 123(11), 2004.
18
Appendix
Figure 1: Estimated approximate productivity shock (vertical axis), materials and capital
0,1
,2,3
,4D
ensi
ty
6 8 10 12 14 16Value Added
Figure 2: Histogram of dependent variable: log(real value added)
19
Table 5: Estimated Productivity Effects of Worker Replacement - OLS results I
Dependent variable: WRR & WRR-sq. - OLS WRR y/n - OLS Empl. > MED - OLS Empl. <= MED - OLS Industries - OLS Firm age > MED - OLS
Real value added Coef. S.E. Coef. S.E. Coef. S.E. Coef. S.E. Coef. S.E. Coef. S.E.
Labour (# of employees) 0.777 (0.021)*** 0.781 (0.022)*** 0.768 (0.026)*** 0.554 (0.050)*** 0.774 (0.021)*** 0.746 (0.028)***
Capital 0.231 (0.015)*** 0.231 (0.015)*** 0.171 (0.016)*** 0.279 (0.024)*** 0.231 (0.015)*** 0.252 (0.022)***
Worker replacement rate -0.198 (0.103)* -0.225 (0.056)*** -0.041 (0.063) -0.089 (0.064) -0.165 (0.078)**
Worker replacement sq. 0.085 (0.095)
Worker replacement y/n -0.043 (0.024)*
WRR # HTM -0.104 (0.132)
WRR # HTS -0.020 (0.147)
WRR # LTM 0.072 (0.165)
WRR # Constr -0.066 (0.121)
Age of firm 0.112 (0.027)*** 0.115 (0.027)*** 0.077 (0.031)** 0.109 (0.042)*** 0.113 (0.027)*** -0.097 (0.086)
Limited liability corporation 0.254 (0.031)*** 0.255 (0.031)*** 0.241 (0.033)*** 0.256 (0.050)*** 0.254 (0.031)*** 0.233 (0.040)***
Managerial experience 0.126 (0.026)*** 0.128 (0.026)*** 0.059 (0.030)** 0.184 (0.040)*** 0.127 (0.026)*** 0.133 (0.036)***
Entrepreneurial experience 0.012 (0.029) 0.011 (0.029) 0.010 (0.031) 0.018 (0.047) 0.011 (0.029) 0.056 (0.038)
Successful entrepreneur
Failed entrepreneur
Highly qual. workers (share) 0.062 (0.045) 0.062 (0.045) 0.094 (0.056)* 0.083 (0.062) 0.064 (0.045) 0.061 (0.057)
Constant / All controls Yes Yes Yes Yes Yes Yes
Observations 5,613 / 0.514 5,613 / 0.514 2,787 / 0.506 2,826 / 0.225 5,613 / 0.514 2,737 / 0.556Notes: *** 1%, ** 5 %, * 10 %. Cluster robust standard errors in parentheses. Additional control variable in all regressions: Funding by the KfW bank.
20
Table 6: Estimated Productivity Effects of Worker Replacement - OLS results II
Dependent variable: Firm age <= MED - OLS Firm age det. - OLS Manag. Exp. - OLS Entrep. Exp. - OLS Entrepr. Exp. (+-) - OLS Qual. Empl. - OLS
Real value added Coef. S.E. Coef. S.E. Coef. S.E. Coef. S.E. Coef. S.E. Coef. S.E.
Labour (# of employees) 0.796 (0.028)*** 0.775 (0.021)*** 0.774 (0.021)*** 0.774 (0.021)*** 0.774 (0.021)*** 0.775 (0.021)***
Capital 0.210 (0.019)*** 0.231 (0.015)*** 0.231 (0.015)*** 0.231 (0.015)*** 0.231 (0.015)*** 0.230 (0.015)***
Worker replacement rate -0.063 (0.057) -0.120 (0.079) -0.110 (0.065)* -0.142 (0.067)** -0.141 (0.066)** -0.080 (0.049)
WRR # Age 3 0.072 (0.105)
WRR # Age 4 0.103 (0.107)
WRR # Age 5 -0.031 (0.177)
WRR # Age 6 -0.201 (0.140)
WRR # Age 7 -0.246 (0.180)
WRR # Age 8 0.201 (0.151)
WRR # Managerial experience 0.011 (0.089)
WRR # Entrepreneurial exp. 0.084 (0.089)
WRR # Successful entrepreneur 0.092 (0.103)
WRR # Failed entrepreneur 0.039 (0.108)
WRR # Hi. qual. work (y/n) -0.072 (0.098)
Age of firm 0.220 (0.047)*** 0.119 (0.029)*** 0.112 (0.027)*** 0.112 (0.027)*** 0.112 (0.027)*** 0.113 (0.027)***
Limited liability corporation 0.259 (0.041)*** 0.253 (0.031)*** 0.254 (0.031)*** 0.255 (0.031)*** 0.252 (0.031)*** 0.255 (0.031)***
Managerial experience 0.122 (0.034)*** 0.127 (0.026)*** 0.126 (0.028)*** 0.127 (0.026)*** 0.133 (0.026)*** 0.127 (0.026)***
Entrepreneurial experience -0.029 (0.039) 0.013 (0.029) 0.012 (0.029) 0.003 (0.031) 0.012 (0.029)
Successful entrepreneur 0.019 (0.037)
Failed entrepreneur -0.013 (0.041)
Highly qual. workers (share) 0.074 (0.065) 0.062 (0.045) 0.063 (0.045) 0.063 (0.045) 0.067 (0.045) 0.070 (0.046)
Constant / All controls Yes Yes Yes Yes Yes Yes
Observations 2,876 / 0.481 5,613 / 0.515 5,613 / 0.514 5,613 / 0.514 5,576 / 0.515 5,613 / 0.514Notes: *** 1%, ** 5 %, * 10 %. Cluster robust standard errors in parentheses. Additional control variable in all regressions: Funding by the KfW bank.
21
Table 7: Estimated Productivity Effects of Worker Replacement - Quits vs. Dismissals - OLS results
Dependent variable: Robust Baseline - OLS Dismissals >= 50 % - OLS Quits > 50% - OLS Quits interacted - OLS Quits (share) interacted - OLS
Real value added Coef. S.E. Coef. S.E. Coef. S.E. Coef. S.E. Coef. S.E.
Capital 0.201 (0.024)*** 0.181 (0.034)*** 0.215 (0.031)*** 0.199 (0.024)*** 0.200 (0.024)***
Labour (# of employees) 0.795 (0.033)*** 0.798 (0.047)*** 0.799 (0.048)*** 0.795 (0.033)*** 0.796 (0.033)***
Worker replacement rate -0.113 (0.060)* 0.011 (0.079) -0.303 (0.092)*** -0.016 (0.077) -0.037 (0.081)
WRR # Quits > 50% -0.269 (0.113)**
WRR # Share of quits -0.201 (0.124)
Share of quits > 50% 0.065 (0.048)
Share of quits 0.033 (0.051)
Age of firm -0.031 (0.044) 0.037 (0.059) -0.125 (0.066)* -0.033 (0.044) -0.033 (0.044)
Limited liability corporation 0.183 (0.049)*** 0.161 (0.071)** 0.193 (0.064)*** 0.182 (0.048)*** 0.182 (0.048)***
Managerial experience 0.129 (0.044)*** 0.130 (0.064)** 0.124 (0.057)** 0.125 (0.044)*** 0.126 (0.044)***
Entrepreneurial experience 0.102 (0.046)** 0.107 (0.063)* 0.113 (0.065)* 0.106 (0.046)** 0.104 (0.046)**
Highly qual. workers (share) 0.064 (0.081) 0.065 (0.123) 0.052 (0.092) 0.058 (0.081) 0.062 (0.081)
GVA p.c. (industry/state) 0.002 (0.003) 0.005 (0.003) -0.002 (0.004) 0.002 (0.003) 0.002 (0.003)
Cutting-edge technology manufacturing -0.147 (0.087)* 0.015 (0.118) -0.302 (0.115)*** -0.142 (0.086) -0.143 (0.086)*
High-technology manufacturing -0.143 (0.099) -0.093 (0.131) -0.204 (0.147) -0.142 (0.099) -0.142 (0.099)
Technology-intensive services -0.016 (0.097) 0.063 (0.136) -0.108 (0.122) -0.010 (0.097) -0.014 (0.097)
Software supply and consultancy -0.200 (0.115)* -0.155 (0.159) -0.271 (0.148)* -0.196 (0.114)* -0.195 (0.114)*
Non-high-tech manufacturing -0.239 (0.093)** -0.103 (0.140) -0.383 (0.121)*** -0.236 (0.093)** -0.236 (0.093)**
Skill-intensive services -0.033 (0.102) 0.078 (0.143) -0.177 (0.131) -0.026 (0.101) -0.029 (0.101)
Other business-oriented services -0.189 (0.110)* -0.175 (0.156) -0.236 (0.148) -0.185 (0.110)* -0.185 (0.110)*
Consumer-oriented services -0.389 (0.085)*** -0.219 (0.124)* -0.590 (0.106)*** -0.384 (0.084)*** -0.385 (0.084)***
Construction -0.058 (0.086) -0.032 (0.114) -0.100 (0.122) -0.057 (0.085) -0.057 (0.085)
Year 2010 0.015 (0.048) 0.038 (0.071) 0.002 (0.066) 0.016 (0.048) 0.015 (0.048)
Year 2011 0.016 (0.049) -0.045 (0.074) 0.124 (0.066)* 0.019 (0.049) 0.017 (0.049)
Year 2012 0.095 (0.050)* 0.065 (0.070) 0.158 (0.076)** 0.097 (0.050)* 0.096 (0.050)*
Constant Yes Yes Yes Yes Yes
Observations / R-squared 1,543 / 0.609 842 / 0.563 701 / 0.676 1,543 / 0.610 1,543 / 0.609Notes: *** 1%, ** 5 %, * 10 %. Cluster robust standard errors in parentheses. Additional control variable in all regressions: Funding by the KfW bank.
22
Table 8: Estimated Productivity Effects of Worker Replacement - Quits vs. Dismissals - LP results
Dependent variable: Robust Baseline LP Dismissals >= 50 % - LP Quits > 50% - LP Quits interacted - LP Quits (share) interacted - LP
Real value added Coef. S.E. Coef. S.E. Coef. S.E. Coef. S.E. Coef. S.E.
Capital 0.172 (0.141) 0.009 (0.222) 0.389 (0.125)*** 0.173 (0.142) 0.172 (0.142)
Labour (# of employees) 0.729 (0.031)*** 0.707 (0.052)*** 0.762 (0.054)*** 0.729 (0.030)*** 0.730 (0.030)***
Worker replacement rate -0.076 (0.055) 0.048 (0.081) -0.261 (0.094)*** 0.022 (0.076) -0.000 (0.077)
WRR # Quits > 50% -0.269 (0.116)**
WRR # Share of quits -0.199 (0.112)*
Share of quits > 50% 0.066 (0.047)
Share of quits 0.034 (0.048)
Age of firm -0.058 (0.042) -0.000 (0.058) -0.129 (0.056)** -0.059 (0.041) -0.059 (0.042)
Limited liability corporation 0.137 (0.048)*** 0.125 (0.078) 0.159 (0.070)** 0.136 (0.047)*** 0.136 (0.048)***
Managerial experience 0.120 (0.041)*** 0.120 (0.072)* 0.120 (0.058)** 0.116 (0.041)*** 0.116 (0.041)***
Entrepreneurial experience 0.096 (0.046)** 0.101 (0.062) 0.092 (0.061) 0.099 (0.045)** 0.097 (0.045)**
Highly qual. workers (share) 0.085 (0.074) 0.069 (0.125) 0.065 (0.089) 0.079 (0.075) 0.083 (0.075)
GVA p.c. (industry/state) 0.002 (0.002) 0.004 (0.004) -0.001 (0.004) 0.002 (0.002) 0.002 (0.002)
Cutting-edge technology manufacturing -0.025 (0.081) 0.164 (0.124) -0.243 (0.131)* -0.020 (0.080) -0.021 (0.080)
High-technology manufacturing -0.109 (0.106) -0.019 (0.134) -0.234 (0.171) -0.107 (0.105) -0.108 (0.105)
Technology-intensive services 0.124 (0.105) 0.227 (0.135)* -0.009 (0.139) 0.131 (0.105) 0.127 (0.105)
Software supply and consultancy 0.011 (0.118) 0.123 (0.178) -0.141 (0.165) 0.015 (0.117) 0.016 (0.117)
Non-high-tech manufacturing -0.144 (0.086)* 0.008 (0.130) -0.330 (0.108)*** -0.141 (0.086) -0.142 (0.086)
Skill-intensive services 0.133 (0.111) 0.255 (0.155)* -0.004 (0.158) 0.141 (0.111) 0.138 (0.111)
Other business-oriented services -0.034 (0.111) 0.041 (0.172) -0.137 (0.155) -0.029 (0.110) -0.030 (0.111)
Consumer-oriented services -0.230 (0.091)** -0.030 (0.128) -0.499 (0.106)*** -0.223 (0.092)** -0.226 (0.091)**
Construction -0.026 (0.084) 0.045 (0.112) -0.133 (0.118) -0.026 (0.083) -0.025 (0.083)
Year 2010 0.018 (0.043) 0.032 (0.067) 0.016 (0.055) 0.019 (0.043) 0.018 (0.043)
Year 2011 0.002 (0.049) -0.058 (0.075) 0.113 (0.063)* 0.005 (0.048) 0.004 (0.049)
Year 2012 0.091 (0.048)* 0.059 (0.065) 0.156 (0.070)** 0.093 (0.048)* 0.091 (0.048)*
Constant Yes Yes Yes Yes Yes
Observations / R-squared 1,543 842 701 1,543 1,543Notes: *** 1%, ** 5 %, * 10 %. Bootstrapped standard errors in parentheses. Additional control variable in all regressions: Funding by the KfW bank.
23
A Worker replacement in German startups
So far, no studies about excess labour turnover in German startups are available. Therefore, a
short descriptive impression is given to facilitate the interpretation of the regression analyses.
The data is taken from KfW/ZEW Start-Up Panel and extrapolated to the population of up to
four year old German start-ups.14 Due to the definition of the worker replacement rate (also
WRR henceforth) in Section 2, calculations can only be done for startups with at least one em-
ployee in the last or the current period.15 In 2012, 31 % of all German startups with employees
show positive labour replacement rates. In these 31 % of the startups, 44% of all employees of
young German firms are employed. Firstly, this shows that excess labour turnover takes place
mainly in larger startups (which have higher potential for labour turnover). Secondly, excess
labour turnover in German startups and its interrelation with firm performance affects a large
proportion of workers employed in young German firms.
Table 9: Worker replacement in 1-4 years old German firms 2012
WRR 2012 Mean MED p75 p90
CTM&HTM 0.07 0.00 0.00 0.18
TIS&Software 0.09 0.00 0.00 0.25
NTM 0.10 0.00 0.07 0.35
Construction 0.19 0.00 0.07 0.75
LTS&Trade 0.16 0.00 0.24 0.50
All 0.15 0.00 0.19 0.50
From literature about excess labour turnover in established firms, large differences in the ex-
tent of replacement hiring can be expected depending on industry and firm age. Table 9 shows
the average and percentiles of the worker replacement rate split for sectors. The average WRR
in German startups in the year 2012 (over all cohorts) is 15 %. Thus, excess labour turnover
in German startups seems to be in a comparable range to the excess turnover rates Haltiwanger
et al. (2012) report for US startups. The average WRR is lowest for high-tech startups in the
manufacturing sector (first row). While the average WRR of high-tech services startups (second
row) and non-high-tech manufacturing startups (third row) is only slightly higher, construction
(fourth row) and low-tech services startups (fifth row) show notably higher WRR. In manu-
facturing as well as in services sectors, excess labour turnover is lower among high-tech firms
compared to non-high-tech firms. In accordance, splitting the results for percentiles of the
WRR distribution shows that in non-high-tech sectors a larger fraction of startups have posi-
tive replacement rates than in high-tech sectors. There are at least two possible explanations
for these sectoral differences. On the one hand, higher replacement rates could be favourable
for firm performance in non-high-tech sectors, since costs for on the job training of employees
might be lower. On the other hand, higher replacement rates in low-tech sectors could reflect
worse bargaining situations of low-tech startups compared to incumbents in their sectors on the
labour market. This could then trigger negative effects on firm performance.
Figure 3 gives an impression of the development of different job and worker flow measures
over the early years of young firms’ lifespan.16 The hiring rate and the resulting job flow and
14Extrapolation is not possible for longer periods due to data restrictions.15All worker flow and job flow measures are calculated as headcounts.16Average values of the foundation cohorts 2005 and 2006 are reported. Due to data restrictions, all values
24
Figure 3: Turnover measures for foundation cohorts 2005 and 2006 over time
worker flow rates are distinctly higher at the beginning of the life-cycle of young firms and
level off after the first five years. This reflects disproportionally fast growth at the beginning
of the firm lifecycle. In contrast, the separation rate and the worker replacement rate show a
more stable pattern. Especially the WRR seems to be very stable over the whole snapshot of the
lifecycle of surviving firms. However, the question if worker replacement is more favourable
or harmful at different stages of the development of a young firm remains. In addition, all
measures for Figure 3 are calculated for firms surviving a seven year period only. Therefore,
the stable labour replacement pattern could reflect the successful personnel strategy of these
firms.
reported in Figure 3 are not extrapolated and only calculated for firms surviving the whole period up to the 7th
year of life.
25